On Fri, Sep 12, 2014 at 05:46:57PM +0100, Catalin Marinas wrote:
> On Fri, Sep 12, 2014 at 05:21:27PM +0100, Daniel Borkmann wrote:
> > On 09/12/2014 06:03 PM, Catalin Marinas wrote:
> > > On Fri, Sep 12, 2014 at 08:11:37AM +0100, Daniel Borkmann wrote:
> > >> Will, Catalin, Dave, this is more or less a heads-up: when net-next and
> > >> arm64-next tree will get both merged into Linus' tree, we will run into
> > >> a 'silent' merge conflict until someone actually runs eBPF JIT on ARM64
> > >> and might notice (I presume) an oops when JIT is freeing bpf_prog. I'd
> > >> assume nobody actually _runs_ linux-next, but not sure about that
> > >> though.
> > >
> > > Some people do.
> > >
> > >> How do we handle this? Would I need to resend this patch when the time
> > >> comes or would you ARM64 guys take care of it automagically? ;)
> > >
> > > I think we could disable BPF for arm64 until -rc1 and re-enable it
> > > together with this patch.
> >
> > Ok, yes, that would mitigate it a bit. Sounds fine to me.
> >
> > > One comment below:
> > >
> > >> --- a/arch/arm64/net/bpf_jit_comp.c
> > >> +++ b/arch/arm64/net/bpf_jit_comp.c
> > > [...]
> > >> +static void jit_fill_hole(void *area, unsigned int size)
> > >> +{
> > >> + /* Insert illegal UND instructions. */
> > >> + u32 *ptr, fill_ins = 0xe7ffffff;
> > >
> > > On arm64 we don't have a guaranteed undefined instruction space (and
> > > Will tells me that on Thumb-2 for the 32-bit arm port it actually is a
> > > valid instruction, it seems that you used the same value).
> >
> > Hm, ok, the boards we've tried out and where Zi tested it too, it worked.
>
> So, if I try this:
>
> $ echo 0xffffffe7 | xxd -r > test.bin
> $ arm-linux-gnueabihf-objdump -m arm -D -b binary test.bin
> ...
> 0: e7ffffff udf #65535 ; 0xffff
...and for Thumb, it ends up as:
0: ffff e7ff vqshl.u64 q15, <illegal reg q15.5>, #63
which does happen to be undefined, but it feels fragile to rely on that
particular instruction form always having UNDEF behaviour.
> Do you use the same constant on arm32?
>
> > > I think the only guaranteed way is to use the BRK #imm instruction but
> > > it requires some changes to the handling code as it is currently used
> > > for kgdb (unless you can use two instructions for filling in which could
> > > generate a NULL pointer access).
> >
> > The trade-off would be that if we align on 8, it would certainly increase
> > the probability to jump to the right offset. Note, on x86_64 we have no
> > alignment requirements, hence 1, and on s390x only alignment of 2.
> >
> > So, on that few (?) boards where UND would be a valid instruction [ as
> > opposed to crash the kernel ], would it translate into a NOP and just
> > 'walk' from there into the JIT image?
>
> On current ARMv8 CPU implementations, the above constant is unallocated
> in the A64 instruction space. But you never know, it may be allocated in
> the future.
>
> I think it's easier if you just use something like BRK #0x100 (opcode
> 0xd4202000) which would trigger a fault in the kernel (kgdb uses #imm
> 0x400 and 0x401).
>
> An unallocated BRK would trigger a fault via do_debug_exception ->
> brk_handler and panic the kernel.
Sounds sensible.
Will
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
